Some tech notes about the @VulkanAPI renderer for #HalfLifeAlyx : volumetric fog is implemented using async compute, per-batch overhead is reduced using bindless texturing (courtesy of my colleague Shanon Drone), and render targets are half size as on DX11
On DX11 we were at times CPU bound on the renderthread, so we implemented multiview instancing to submit to a double-wide target. On Vulkan we stuck with dual command buffer submission for each eye, which while not always as efficient on the GPU helps our texture streamer out by
reducing GPU memory use overall (our render targets are HUGE, especially with headsets like the Index). The bindless path helps a lot with per-batch overhead, it was complex to fit into the engine, but basically we got down to a push constant update and VB/IB bind between draws
I'm really pleased with how the Linux version runs on AMD RADV where we have async reprojection in SteamVR via a high priority compute queue. We look forward to getting that on NVIDIA but are awaiting driver support, so for now we recommend mesa/AMD.
